The document discusses telco cloud and network virtualization technologies including NFV and SDN. It provides an overview of how NFV and SDN enable programmability and virtualization of network resources to provide flexibility. NFV allows network functions to run in software on commercial off-the-shelf hardware, while SDN separates the network control and forwarding planes to enable centralized programmable network control. Together NFV and SDN can optimize resource utilization and simplify network management.

1. Telco Cloud and
Network (NFV and
SDN)

2. 2

3. Global Mobile Traffic
3
(Compound Annual Growth Rate)

4. Driving Forces
4

5. 5
5G Use Cases
• Massive-Machine Type Communication (mMTC)
• Mission Critical-Machine Type Communication (MC-MTC)

6. Diverse Requirements of IoT
6

7. One Network – Multiple Industries
7

8. 8

9. Telco Cloud Networks
 SDN/NFV enable programmability & Cloud enable virtualization of
network resources
 High level of flexibility and programmability in individual domains
(mobile core, radio access network and transport network).
 Cross-domain programmability and orchestration.
9

10. Telco Network
10
Change the business
model of Telecom
Biggest technological
revolution
Innovative more
quickly
Change the way
network are deployed
Change the way
consumers are enabling
services on the fly

11. Why Cloud Network are required
in Telco?
11

12. Cloud Network?
12

13. What is a Telco Cloud?
 A Telco Grade (aka Carrier Grade) Cloud will be a cloud
that can support telco grade applications
 Telco Grade requirements:
 High availability
 High performance (large number of transactions, scalability)
 Serviceability
 Long life time
 Security
 Real-time behavior
 Standard-compliant HW
13

14. Why Telco Clouds?
14
Problem of Telco Operators

15. Get Right Capacity at Right Time
15

16. Get Right Capacity at Right Time
16

17. Get Right Capacity at Right Time
17

18. Why Telco Cloud?
18
Problem of Telco Operators

19. How Telco Cloud Helps?
19

20. Solution!
20

21. Migration (Traditional to Cloud)
21

22. NFV & SDN
22

23. NFV & SDN
23

24. NFV Initiative
24

25. Introduction to NFV
25

26. Introduction to NFV
26

27. Introduction to NFV
27

28. Traditional Network Functions (NFs)
 Proprietary devices/boxes for different NFs
 Network services rely on different type of appliances
 New services into today’s network is becoming increasingly difficult
due to:
 Proprietary nature of appliances
 Diverse and purpose built hardware
 Cost (increase CapEx & OpEx)
 Short life cycle of appliances
 Lack of space
 Energy for middle-boxes
 Lack of skilled professionals to integrate services
 Recently, NFV proposed to alleviate these problems

29. Introduction– Network Function
Virtualization
 NFV proposed by ESTI, an Industry Specification Group (ISG)
 Allows to implement NFs (Network Functions) in software.
 Virtualized the NFs carried by proprietary HW.
 Decoupling NFs from underlying appliances
 NFs can runs on commodity hardware (i.e. servers, storage, switches)
 Accelerate deployment of new services and NFs

30. Network Function Virtualization
30

31. Why We Need NFV?
 Virtualization: Use network resource without worrying about where it is
physically located, how much it is, how it is organized, etc.
 Orchestration: Manage thousands of devices
 Programmable: Should be able to change behavior on the fly.
 Dynamic Scaling: Should be able to change size, quantity
 Visibility: Monitor resources, connectivity
 Performance: Optimize network device utilization
 Multi-tenancy
 Service Integration
 Openness: Full choice of Modular plug-ins

32. NFV Architecture
32

33. NFV Architecture
33

34. NFV Architecture
34

35. NFV Architecture
35

36. NFV Architecture
36

37. Why SDN?
37

38. NFV vs SDN
38

39. What is SDN?
39

40. Traditional Network Devices
 Typical Networking Software
 Management plane
 Control Plane – The brain/decision maker
 Data Plane – Packet forwarder

41. Traditional Network Devices
41

42. What is SDN?

43. What is SDN?
43

44. Basic concepts of SDN
 Separate Control plane and Data plane entities.
 Network intelligence and state are logically centralized.
 The underlying network infrastructure is abstracted from
the applications.
 Execute or run Control plane software on general
purpose hardware.
 Decouple from specific networking hardware.
 Use commodity servers and switches.
 Have programmable data planes.
 Maintain, control and program data plane state from a
central entity.
 An architecture to control not just a networking
device but an entire network.

45. SDN Framework
45

46. SDN Architecture
 Application Layer:
 Focusing network services
 SW apps communicating with the
control layer
 Control-Plane Layer:
 Core of SDN
 Consists of a centralized controller
 Logically maintains a global and
dynamic view
 Take requests from application
layer
 Manage network devices via
standard protocols
 Data-Plane Layer:
 Programmable devices
 Support standard interfaces

47. Data Plane – Forwarding Devices

48. Southbound Interface – OpenFlow
 Forwarding elements are controlled
by an open interface
 OpenFlow has strong support from
industry, research and academia
 OpenFlow, standardized information
exchange b/w two planes
 Provides controller-switch
interactions
 OpenFlow 1.3.0 provides secure
communication using TLS
 Communicate with controller via
OpenFlow protocol

49. Control Plane – SDN Controller
 Controller provides a programmatic
interface to the network
 Give a logically centralized global
view of network
 Simplification of policy enforcements
and management
 Some concerns:
 Control scalability
 Centralized vs. Distributed
 Reactive vs. Proactive Policies

50. Control Plane – SDN Controller (Cont..)

51. Northbound Interfaces & SDN Applications
 Northbound Interface:
 A communication interface between Control plane and
Applications
 There is no currently accepted standard for northbound
interfaces
 Implemented on ad hoc basis for particular application
 SDN Applications:
 Traffic Engineering
 Security
 QoS
 Routing
 Switching
 Virtualization
 Monitoring
 Load Balancing
 New Innovations???

52. NFV vs SDN
52

53. SDN & NFV Relationship
 Concept of NFV originated from SDN
 NFV and SDN are complementary.
 One does not depend upon the other.
 You can do SDN only, NFV only, or SDN and NFV
 Both have similar goals but approaches are very different
 SDN needs new interfaces, control modules, applications
 NFV requires moving network applications from dedicated
hardware to virtual containers on commercial-off-the-shelf
(COTS) hardware

54. NFV vs. SDN
 NFV can serve SDN by
virtualizing SDN controller
 Implements NFs in software
 Reducing CapEx, OpEx, space,
power consumption .
 Decouple network function from
proprietary hardware and
achieve agile provisioning and
deployment
 SDN serves NFV by providing
programmable connectivity b/w
VNFs to optimize traffic
engineering & steering
 Central control and
programmable architecture for
better connectivity
 Network abstractions to enable
flexible network control,
configuration and innovation
 Decouple control plane from
data plane forwarding to provide
a centralized controller

55. Software-Defined NFV Architecture

56. Software-Defined NFV–Service Chaining

57. NFV
Virtualize
CLOUD
Scale
SDN
Control
Management & Orchestration
Cross Domain Control, Orchestration & Management
SDN,Cloudand NFV

58. Introduction of ONOS
 ONOS (Open Network Operating System) is an open
source SDN OS
 Developed in concert with leading SPs, vendors, R&E
network operators and collaborators
 Specially targeted: Service Providers and mission critical
networks.
 ONOS main goals:
 Liberate network application developers from knowing the details
of proprietary hardware
 Free from the operational complexities of proprietary interfaces
and protocols
 Re-enable innovation to happen for both network hardware and
software

59. Why ONOS??
 Several open source controllers (NOX, Beacon, SNAC,
POX, etc)
 ONOS will:
 Bring carrier grade network (scale, availability, and performance)
to SDN control plane
 Enable web style agility
 Help SPs to migrate existing networks
 Lower SP CapEx & OpEx

60. ONOS Vision for SPs Networks
 Enabling SP SDN adoption for carrier-grade service and network innovation

61. Key Elements of ONOS
 Modular, Scalable, Resilient with Abstractions

62. ONOS Architecture

63. Defining Features of ONOS
 Distributed Core:
 Provides scalability, high availability, and performance
 Bring carrier grade features
 Run as a cluster is one way that ONOS brings web style agility
 Northbound abstraction/APIs:
 Include network graph and application intents to ease development
of control, management, and configuration services
 Southbound abstraction/APIs:
 Enable pluggable southbound protocols for controlling both
OpenFlow and Legacy devices
 A key enabler for migration from legacy devices to OpenFlow-based
white boxes
 Software Modularity:
 Easy to develop, debug, maintain, and upgrade ONOS as a
software system

64. Distributed ONOS Architecture

65. Software Modularity
 ONOS software is easy to enhance, change, and maintain
 Great care into modularity to make it easy for developers
 At the macro level the Northbound and Southbound APIs provide an initial
basis for insulating Applications, Core and Adapters from each other
 New applications or new protocol adapters can be added as needed without
each needing to know about the other
 Rely heavily on interfaces to serve as contracts for interactions between
different parts of the core

66. ONOS Initial SPs Use Cases

67. SONA (Simplified Overlay Network Architecture)
 ONOS applications
 Provides OpenStack Neutron ML2 mechanism driver
and L3 service plugin
 Optimized tenant network virtualization service
 Provisioning an isolated virtual tenant network uses
VXLAN based L2 tunneling or GRE/GENEVE based L3
tunneling with OpenvSwitch (OVS)
 Horizontal scalability of a gateway node
67

68. SONA (Simplified Overlay Network Architecture)
68

69. Example of SONA
69

70. Introduction to High
Performance Computing
(HPC) & High Throughput
Computing (HTC)

71. 71
Anatomy of a Computer

72. 72
Multicore
 Cores share path to
memory
 SIMD instructions +
multicore make this an
increasing bottleneck!

73. Performance
 The performance (time to solution) on a single computer
can depend on:
 Clock speed – How fast the processor is
 Floating point unit -- how many operands can be operated on
and what operations can be performed?
 Memory latency – what is the delay in accessing data?
 Memory bandwidth – how fast can we stream data from memory
 I/O to storage – how quickly can we access files?
 For parallel computing you can also be limited by the
performance of the interconnection
73

74. Performance (Cont.)
 Application performance often described as:
 Compute bound
 Memory bound
 IO bound
 Communication bound
 For computational science
 Most calculations are limited by memory bandwidth
 Processor faster than memory access
74

75. 75
HPC Architectures
 All Cores have the same access to memory, e.g. a multicore laptop.
Symmetric Multi-processing

76. 76
HPC Architectures (Cont.)
Distributed Memory Architecture

77. 77
HPC Architectures (Cont.)
 In a real system:
 Each node will be a
shared-memory
system
 E.g. multicore processor
 The network will have
some specific topology
 E.g. a regular grid
Distributed/Shared Memory Hybrids

78. The Flood of Data
78
Why HPC?

79. Why HPC? (Cont.)
79

80. Why HPC? (Cont.)
 Big data and scientific simulations need greater
computer power.
 Single-core processors can not be made that
have enough resources for the simulations
needed.
 Making processors with faster clock speeds is difficult
due to cost
 Expensive to put huge memory on a single processor
 Solution parallel computing – divide up the work
among numerous linked systems
80

81. Generic Parallel Machines
81
 Good conceptual model is collection of multicore
laptops.
 Connected together by a network
 Each laptop is called a
compute node
 Each has its own OS and
network
 Suppose each node is
quad core
 Total system has 20
processor-cores

82. Parallel Computing?
 Parallel computing and HPC are intimately
related
 Higher performance requires more processor-cores
 Understanding of different parallel programming
models allows you to understand how to use
HPC resources efficiently
 Also allow you to better understand and critique
work that uses HPC in your research area
82

83. What is HPC?
 Leveraging distributed compute resources to solve
complex problems with large datasets
 Terabytes to petabytes to zettabytes of data
 Results in minutes to hours instead of days or weeks
83

84. Differences from Desktop Computing
 Do not log on to compute nodes directly
 Submit jobs via batch scheduling systems
 Not a GUI-based environment
 Share the system with many users
 Resources more tightly monitored and controlled
 Disk quotas
 CPU usage
84

85. Typical HPC System Layout
85

86. 86
Typical Software Usage Flow

87. HPC Applications?
87

88. High Throughput
Computing(with HTCondor)
88

89. Serial Computing
 What many programs look
like:
 Serial execution, running on one
processor (CPU core) at a time
 Overall compute time grows
significantly as individual tasks
get more complicated (long) or if
the number of tasks increases
 How can you speed things up?
89

90. High Throughput Computing (HTC)?
 Parallelize!
 Independent tasks run on different cores
90

91. High Performance Computing (HPC)?
91

92. High Performance Computing (HPC)?
 Benefits greatly from:
 CPU speed + homogeneity
 Shared filesystems
 Fast, expensive networking (e.g.
InfiniBand) and servers co-located
 Scheduling: Must wait until all processors
are available, at the same time and for
the full duration
 Requires special programming (MP/MPI)
 What happens if one core or server fails
or runs slower than the others?
92

93. High Throughput Computing (HTC)?
 Scheduling: only need 1 CPU core for each (shorter
wait)
 Easier recovery from failure
 No special programming required
 Number of concurrently running jobs is more important
 CPU speed and homogeneity are less important
93

94. High Throughput vs High Performance
 HTC
 Focus: Large workflows of
numerous, relatively
small, and independent
compute tasks
 More important: maximized
number of running tasks
 Less important: CPU speed,
homogeneity
94
 HPC
 Focus: Large workflows of
highly-interdependent sub-
tasks
 More important: persistent
access to the fastest cores,
CPU homogeneity, special
coding, shared filesystems,
fast networks

95. Example..
 You need to process 48 brain images for each of 168
patients. Each image takes ~1 hour of compute time.
 168 patients x 48 images = ~8000 tasks = ~8000 hrs
95

96. Distributed Computing
 Use many computers, each running one instance of
our program
 Example:
 1 laptop (1 core) => 4,000 hours = ~½ year
 1 server (~20 cores) => 500 hours = ~3 weeks
 1 large job (400 cores) => 20 hours = ~1 day
 A whole cluster (8,000 cores) = ~8 hours
96

97. Break Up to Scale Up
 Computing tasks that are easy to break up are easy
to scale up.
 To truly grow your computing capabilities, you also
need a system appropriate for your computing task!
97

98. What computing resources are available?
 A single computer?
 A local cluster?
 Consider: What kind of cluster is it? Typical clusters tuned for
HPC (large MPI) jobs typically may not be best for HTC
workflows! Do you need even more than that?
 Open Science Grid (OSG)
 Other
 European Grid Infrastructure
 Other national and regional grids
 Commercial cloud systems (e.g. HTCondoron Amazon)
98

99. Example Local Cluster
 UW-Madison’s Center for High Throughput
Computing (CHTC)
 Recent CPU hours:
 ~130 million hrs/year (~15k cores)
 ~10,000 per user, per day (~400 cores in use)
99

100. Open Science Grid (OSG)
 HTC for Everyone
 ~100 contributors
 Past year:
 >420 million jobs
 >1.5 billion CPU hours
 >200 petabytes transferred
 Can submit jobs locally, they backfill across
the country-interrupted at any time (but not
too frequent)
 http://www.opensciencegrid.org/
100